Aimed at the problem that air target group intent is often difficult to be identified rapidly under condition of imbalanced and difficult classification, an intent recognition method is proposed based on moving-window estimation of the temporal convolution self-attention network model. First, the proposed model is intended to preprocess the feature data by the moving-window estimation method. Second, the flow information of multi-dimensional time series feature data is quickly extracted by the temporal convolution network (TCN). Finally, the self-attention mechanism is used to capture the key features from each feature datum and optimize the weights. The simulation results show that this method improves the training efficiency and classification accuracy for the intent recognition of hard-sample in imbalanced samples.